This dataset is a compilation of ClinicalTrials.gov data (a U.S. Government
database) by the dataset author. The compilation is licensed under
CC-BY-NC-4.0. As a condition of accessing these files, you agree to:

Cite this dataset, and attribute the underlying data to ClinicalTrials.gov,
in any publication or derivative work.
Not use these files, or copies of them, for commercial purposes without a
separate commercial license from the author. Examples: ML training data
for commercial models, redistribution in a commercial product or service,
inclusion in databases offered for sale, or quantitative trading /
commercial market analysis.

These terms apply to this dataset only, not to data you obtain directly from
ClinicalTrials.gov. For commercial licensing, contact licensing@anamnesisdata.com
or visit https://anamnesisdata.com.

Clinical Trials Version History

The Clinical Trials Version History dataset by Anamnesis Data: structured medical and scientific data for biotech, pharma and healthcare research. Public trial registries show only a trial's latest state. This dataset keeps every version, so you can see what a trial said on any past date, or when a sponsor moved a completion date.

606,620 trials (4,456,564 versions), one folder of Parquet files per table (79 tables, 541,236,196 rows). Access is gated; request it above. A free sample of 20 complete trials has the same tables and columns.

Example queries

sample_queries.sql in the sample holds ready-to-run DuckDB queries, from a trial's current record to serious adverse event rates. They run unchanged against this dataset's setup.sql, which creates one view per table; on the full data, add a LIMIT to the row-level ones.

Quick start

Download the repository, then in DuckDB, from that folder:

.read setup.sql          -- one view per table, named after its folder

-- the current public record of every trial
SELECT nct_id, nct_version, overall_status, lead_sponsor_name
FROM core
WHERE last_update_post_date IS NOT NULL      -- leaves out versions that failed review
QUALIFY row_number() OVER (PARTITION BY nct_id ORDER BY nct_version DESC) = 1;

Or read a table in place with a token from an account that has been granted access:

CREATE SECRET hf_token (TYPE huggingface, PROVIDER credential_chain);  -- uses `hf auth login`
SELECT * FROM 'hf://datasets/anamnesis-data/clinical_trials_history/core/*.parquet' LIMIT 5;

from datasets import load_dataset
core = load_dataset("anamnesis-data/clinical_trials_history", "core", split="train")

What's inside

Each folder is one table. core has one row per trial version; interventions, conditions, phases, arm_groups, locations and the results_* tables (participant flow, baseline, outcomes, adverse events) hang off it. Tables with nct_version join to core on (nct_id, nct_version); per-trial tables join on nct_id.

Things to know:

Versions. nct_version is the source's own 0-based version number for a trial.
Public dates. A version is public from last_update_post_date. Versions that failed the registry's review were never public: their last_update_post_date is empty, and reset_notice_post_date holds the date of the review's reset notice.
No "current" flag. The current record is a trial's highest public nct_version, as in the query above. It is the latest version held here, which can trail the live record.
Versions as captured. Each version is kept as captured; the registry later re-derives some fields in about 1 trial in 10. version_history follows the registry's current change list.
Obsolete IDs. Every table ends with is_obsolete and obsolete_by, which flag registrations since merged into another (22 trials). Filter on NOT is_obsolete to drop them.
List entries. List tables (for example interventions, arm_groups, secondary_ids, overall_officials, conditions and keywords) keep every source entry, in order: ord is its position in the source list, so entries that repeat a name are separate rows.
Text. No field holds HTML: entities are decoded (A & B becomes A & B), tags are removed, and <word> used as brackets becomes (word). Rich-text fields (descriptions, eligibility criteria, results comments) are Markdown, like the public ClinicalTrials.gov API.

Releases

Each release is tagged vYYYY.MM.DD; pass it as revision= (or @vYYYY.MM.DD in an hf:// path) to pin one. A tag moves if the same day is republished; pin a commit hash for a fixed snapshot. This release's newest change: 2026-10-09.

Breaking changes since `v2026.09.28`

Code written against v2026.09.28 needs these changes; that release stays available under its tag.

interventions is keyed by (nct_id, nct_version, ord): entries with the same name and type are separate rows, and entries without a name or type are kept.
New ord column, ordered by it: arm_groups, collaborators, overall_officials, secondary_ids, conditions, keywords, phases, who_masked, std_ages, ipd_info_types, id_aliases. intervention_other_names gains intervention_ord and other_name_ord.
intervention_arm_links has one row per source element (source, parent_ord, item_ord, raw_label, intervention_ord, arm_group_ord); intervention_type uses the enum values.
core.last_update_post_date is empty for versions that failed review; the new last column reset_notice_post_date holds their reset-notice date.
Plain-text fields are decoded (see Text); version_history.module_labels is {} rather than NULL for changes that list no module.
No HTML in text: rich-text fields are Markdown, and tags and entities are removed (see Text).

Repository move

Releases up to v2026.09.26 were published at brbk/clinical_trials_history; new releases are published here only. Access approvals do not carry over, so request access again here.

Source and license

Trial records come from ClinicalTrials.gov, a U.S. Government database operated by the National Library of Medicine at NIH; no proprietary rights are claimed in them, and the site's terms apply to them. We restructured the records into tables and a subset of fields; dates given only as a month or year are set to its first day (a *_precision column, where present, keeps the original), and HTML is removed from text and rich-text fields are converted to Markdown. Other values are as returned.

The compilation (structure, schema, version history and documentation) is licensed CC BY-NC 4.0. Access is gated: you agree to cite this dataset and attribute the data to ClinicalTrials.gov, and not to use these files commercially without a separate license. These terms do not apply to data you obtain directly from ClinicalTrials.gov. For commercial licensing, contact licensing@anamnesisdata.com or visit anamnesisdata.com.

Citation

@misc{clinical_trials_history_2026,
  title  = {Clinical Trials Version History},
  author = {{Anamnesis Data}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/anamnesis-data/clinical_trials_history}
}

Downloads last month: 72

Total file size:

2.86 GB

You need to agree to share your contact information to access this dataset